Redact PII from a document

Every entity GLiNER returns carries the page boxes its characters fall in, which is what makes redaction a drawing operation rather than a search-and-replace.

import { Pipeline, ImageDrawBoxes, configure } from '@stabrise/scaledp'
import { PdfToImage } from '@stabrise/scaledp/pdf'
import { PaddleTextRecognizer } from '@stabrise/scaledp/ocr'
import { GlinerNer } from '@stabrise/scaledp/ner'

const pipeline = new Pipeline([
    new PdfToImage({ resolution: 300 }),
    new PaddleTextRecognizer({ keepFormatting: true }),
    new GlinerNer({
        inputCol: 'text',
        labels: ['person', 'email', 'phone', 'address', 'credit_card_number'],
        threshold: 0.5,
    }),
    new ImageDrawBoxes({
        inputCols: ['image', 'ner'],
        outputCol: 'redacted',
        filled: true,
        color: '#000000',
        padding: 2,
    }),
])

const rows = await pipeline.transform(file)

filled: true paints the box rather than outlining it. padding: 2 covers antialiased edges that a tight box leaves visible.

Reviewing before redacting

Redaction is not something to do on trust. Draw the entities in a visible colour first, with their labels, and let a human confirm:

new ImageDrawBoxes({
    inputCols: ['image', 'ner'],
    outputCol: 'review',
    color: '#ff5c8a',
    lineWidth: 2,
    displayDataList: ['entity_group', 'score'],
})

import { renderInto, showNer, visualizeNer } from '@stabrise/scaledp/display'

renderInto('#entities', showNer(row.ner, { limit: 0 }))
renderInto('#inline', visualizeNer(row.text, row.ner))

Narrowing what is redacted

Two different knobs:

• labels changes the inference. Removing 'address' means the model is not asked about addresses at all.
• whiteList on GlinerNer, or whiteList / blackList on ImageDrawBoxes, filters afterwards. Use these to run one inference and paint several different redaction sets from it.

Raising threshold trades recall for precision. For redaction, recall is usually what matters — a missed phone number is worse than a redacted false positive — so a lower threshold plus human review is often the right shape.

Precision of the boxes

The redaction is only as tight as the boxes underneath it. Line-level detectors give line-level redaction, which over-covers. For word-level boxes:

• TesseractOcr returns word boxes directly, or
• TesseractRecognizer with boxLevel: 'word' after a detector — see Detect, then recognise.

The redacted page

row.redacted is a ScaleDpImage — encoded bytes plus dimensions. Write it out, or hand it to showImage. Nothing left the browser at any point.